6G-oriented cross-domain knowledge-driven network intelligent scheduling system and method

Abstract

The invention discloses a 6G-oriented cross-domain knowledge-driven network intelligent scheduling system and a 6G-oriented cross-domain knowledge-driven network intelligent scheduling method. The architecture comprises a cross-domain knowledge collaborative evolution module, a network closed-loop management module and a multi-normal-form learning decision module, the cross-domain knowledge collaborative evolution module is used for disassembling a network state into an environment domain, a network domain and a user behavior domain, and online incremental fusion of three-domain knowledge is realized through a graph neural network; dynamically updating the global knowledge base based on knowledge distillation; the network closed-loop management module is used for constructing a function closed loop of perception-reasoning-knowledge generation-decision issuing-verification optimization-memory retrieval; and the multi-normal-form learning decision module is used for fusing meta learning, reinforcement learning, federal learning and self-supervised learning to realize rapid migration and continuous evolution of the strategy. According to the method, the network data throughput, the fault recovery speed, the flow prediction precision, the resource fairness, the path efficiency and the task success rate are improved, and the method is suitable for complex 6G scenes such as air-ground integration and low-altitude traffic control.

Description

Technical Field

[0001] This invention relates to the field of intelligent management and control technology for mobile communication networks, and in particular to a cross-domain knowledge-driven intelligent network scheduling system and method for 6G. Background Technology

[0002] With the rapid evolution of mobile communication technology, 6G networks have been given the core mission of supporting ultra-high bandwidth, ultra-low latency, massive connectivity, and intelligent collaboration across the entire domain. Their application scenarios cover complex fields such as integrated space-ground communication, low-altitude traffic control, autonomous driving, and the industrial internet. These scenarios place unprecedentedly stringent demands on the network's real-time performance, dynamic adaptability, and global optimization capabilities, making traditional network management and maintenance technologies insufficient to meet these needs.

[0003] Currently, network automation operation and maintenance technology has evolved from Software-Defined Networking (SDN) to Knowledge-Defined Networking (KDN). SDN achieves centralized network management by separating the control plane and data plane, giving the controller global visualization capabilities and optimizing routing decisions to some extent. However, facing the hundreds of billions of terminal accesses, dynamic heterogeneous topology switching, and diverse service scenarios in 6G networks, SDN's static configuration and local optimization mechanisms cannot meet the needs of real-time automated operation and maintenance across the entire network. To address this, KDN introduces a "knowledge plane" on top of the SDN architecture, using machine learning technology to extract actionable knowledge from massive amounts of network data to support real-time intelligent management and adaptive optimization. However, its core limitation lies in the fact that knowledge extraction relies on single-domain offline learning, making it difficult to cope with cross-domain dynamic scenarios.

[0004] At the level of network optimization methods, early research mostly adopted combinatorial optimization or integer programming models, simplifying problems such as resource allocation and routing scheduling into single objectives such as minimizing link load, and solving for approximate optimal solutions through mathematical programming. However, such methods are highly sensitive to network scale and dynamics, and have poor adaptability in 6G dynamic scenarios such as drone swarms and autonomous driving; while stochastic combinatorial optimization models can cope with some uncertainties, they cannot meet real-time requirements due to slow iterative convergence.

[0005] In recent years, Deep Reinforcement Learning (DRL) has become a cutting-edge technology for automated network management due to its ability to learn autonomously from real-time data without relying on predefined models. Graph Neural Network (GNN)-based DRL architectures have shown potential in combinatorial optimization scenarios, but their generalization ability is limited by the training environment. Retraining is required in new topologies or cross-domain fusion scenarios, making it difficult to adapt to the dynamic environment of 6G. Federated learning, through distributed node collaborative training and sharing only model parameters, improves global decision-making capabilities while protecting data privacy, but lacks a cross-domain knowledge collaboration mechanism, making rapid policy transfer difficult.

[0006] Overall, existing technologies face three major bottlenecks:

[0007] 1. Insufficient local optimization capabilities: Existing solutions (such as OSPF and MPDRL) mostly focus on a single network domain or service domain, rely on traffic-related indicators, and ignore the combined impact of environmental domain (such as electromagnetic interference and weather changes) and user behavior domain (such as service level agreements). This results in a single decision-making dimension, making it difficult to meet the multi-domain collaborative needs of complex scenarios such as 6G integrated terrestrial and aerial networks and low-altitude traffic.

[0008] 2. Weak generalization ability: The trained model is difficult to adapt to unknown topologies, new business scenarios, or dynamic network states, lacking cross-scenario migration and rapid deployment capabilities, and failing to meet the "plug and play" requirements of 6G. For example, although the Meta-UAV framework introduces a meta-learning mechanism, it only relies on environmental situation awareness and does not combine network state and user intent for cross-domain joint optimization, resulting in low policy transfer efficiency.

[0009] 3. Single optimization objective: Existing frameworks often focus on a single optimization indicator, making it difficult to take into account multi-dimensional needs and achieve a balance between overall network evolution and multi-level service quality assurance. Summary of the Invention

[0010] The purpose of this invention is to provide a cross-domain knowledge-driven intelligent scheduling system and method for 6G networks that features strong real-time performance, strong dynamic adaptability, and strong global optimization capabilities.

[0011] The technical solution to achieve the purpose of this invention is: a cross-domain knowledge-driven intelligent network scheduling system for 6G, including a cross-domain knowledge collaborative evolution module, a network closed-loop management module, and a multi-paradigm learning decision module;

[0012] The cross-domain knowledge collaborative evolution module is used to decompose the network state into the environment domain, network domain and user behavior domain, realize the online incremental fusion of knowledge in the three domains through graph neural network, and dynamically update the global knowledge base based on knowledge distillation.

[0013] The network closed-loop management module is used to construct a functional closed loop of perception-reasoning-knowledge generation-decision issuance-verification optimization-memory retrieval;

[0014] The multi-paradigm learning decision module is used to integrate meta-learning, reinforcement learning, federated learning, and self-supervised learning to achieve rapid policy transfer and continuous evolution.

[0015] Furthermore, the cross-domain knowledge collaborative evolution module includes a domain splitting unit, an online fusion unit, and a knowledge base updating unit;

[0016] The domain splitting unit is used to define environmental domain data as meteorological parameters, electromagnetic characteristics and geographic information, network domain data as topology, resource status and performance indicators, and user behavior domain data as business type, service requirements and task priority.

[0017] The online fusion unit uses a graph neural network to uniformly encode the heterogeneous data across the three domains, and dynamically allocates weights for each domain through an attention mechanism. These attention weights are calculated using the Softmax function.

[0018]

[0019] in, For domain Attention weights , For domain ,domain eigenvectors, To adjust the parameters;

[0020] The knowledge base update unit compresses new and old knowledge through knowledge distillation, and defines a distillation loss function. :

[0021]

[0022] in, For cross-entropy loss, Let KL divergence be the KL divergence. This is the balance coefficient.

[0023] Furthermore, the environmental domain data includes air density, wind speed, rainfall, channel quality, interference source distribution, coverage blind spots, and path attenuation; the network domain data includes multi-hop topology, air-to-ground handover delay, network slice resource allocation, link utilization, SNR, BER, and latency jitter; and the user behavior domain data includes traffic characteristics, flight paths, service level requirements, and task priorities.

[0024] Furthermore, the graph neural network employs a knowledge graph embedding method based on TransE, using knowledge graph triples. Embedded target function for:

[0025]

[0026] in, , , These are head entities, relations, and tail entities, where an entity refers to the subject in the knowledge graph, a relation refers to the semantic relationship between entities, and a tail entity refers to the object in the knowledge graph. These are the embedding vectors for the head entity, relation, and tail entity, respectively. For the set of positive sample triples, For negative sample set, This is a marginal parameter used to control the spacing between positive and negative samples.

[0027] Furthermore, the network closed-loop management module includes a multi-domain perception layer, an intent parsing layer, a decision execution layer, and a verification feedback layer;

[0028] The multi-domain perception layer is used to deploy distributed sensors to collect data from three domains and output multimodal feature vectors in a unified format:

[0029]

[0030] in For time steps; The unified eigenvector representing time t; Indicates the characteristics of the environmental domain; Represents network domain characteristics; Represents user behavior domain characteristics;

[0031] The intent parsing layer uses a lightweight NLP model to convert natural language instructions into structured intents. :

[0032]

[0033] in For maximum delay, For minimum bandwidth, To circumvent the strategy, Prioritize the scene;

[0034] The decision execution layer uses a Graph Convolutional Network (GCN) to propagate cross-domain knowledge and generate a globally optimal policy. The feature update formula for the GCN is as follows:

[0035]

[0036] in, For nodes In the Features of the layer For the set of neighboring nodes, For the first The learnable parameter matrix of the layer, Here, j represents the activation function; j represents the index of the neighboring node.

[0037] The verification feedback layer monitors throughput, packet loss rate, latency fluctuations, and energy consumption in real time, and adjusts the strategy through reinforcement learning feedback. The reward update formula for reinforcement learning is as follows:

[0038]

[0039] in, This represents the time difference error, used for policy updates; For instant rewards, As a discount factor, The current state value function, Estimate the value function for the next state.

[0040] Furthermore, the intent parsing layer also includes an experiential knowledge retrieval subunit, used to calculate the semantic similarity between the current intent vector and the experiential knowledge triples. :

[0041]

[0042] in, This is the current intent vector. For the first A strategy template, when When the value exceeds the threshold, the recommended activation strategy template is provided.

[0043] Furthermore, the verification feedback layer will perform feedback injection into the knowledge graph to form new triples. :

[0044]

[0045] in , For entities, To achieve a closed-loop evolution of intention-experience-knowledge in the relationships between entities.

[0046] Furthermore, the multi-paradigm learning decision module includes federated learning units, meta-learning units, and reinforcement learning units;

[0047] The federated learning unit is used to aggregate the local model parameters of distributed nodes, and the global model parameter update formula is:

[0048]

[0049] in, These are global model parameters. For nodes Local parameters, For nodes The amount of data, Total data volume;

[0050] The meta-learning unit optimizes meta-parameters through inner and outer loop training, with the inner loop fine-tuning task-specific parameters:

[0051]

[0052] Outer loop updates meta parameters:

[0053]

[0054] in , These are the learning rates for the inner and outer loops, respectively. The loss function; For the task Dedicated model parameters; These are the meta-parameters, i.e., the initial shared parameters; For the task In parameters The following losses; For the task Parameters after fine-tuning The following losses; For parameters The gradient; For meta-parameters The gradient;

[0055] The reinforcement learning unit employs a deep reinforcement learning optimization strategy, with the objective function being:

[0056]

[0057] in, As a strategy, For trajectory, For strategy Expected return; For a moment Discount factor; For a moment Instant rewards; In strategy Lower trajectory The expectation.

[0058] Furthermore, the federated learning unit also introduces node performance weights, the weights being calculated using the following formula:

[0059]

[0060] in, For nodes At any moment Aggregate weights; For nodes Comprehensive performance indicators; For other nodes Performance metrics; The total number of nodes participating in federated learning; T is the temperature factor, which controls the sensitivity of the control weights.

[0061] A cross-domain knowledge-driven intelligent scheduling method for 6G is disclosed. This method is based on the aforementioned cross-domain knowledge-driven intelligent scheduling system for 6G and is applied to low-altitude traffic control scenarios. It includes cross-domain networking, intent-guided agent management, and swarm path planning, and optimizes throughput, fault recovery time, traffic prediction accuracy, decision latency, resource fairness, path efficiency, and task success rate.

[0062] Compared with the prior art, the significant advantages of this invention are:

[0063] (1) The knowledge required for network intelligence is divided into three categories: environment domain, network domain, and user behavior domain. Online incremental fusion is achieved based on graph neural network. New and old knowledge are compressed and updated through knowledge distillation technology to continuously improve the global knowledge base and ensure the coordinated evolution of knowledge in each domain.

[0064] (2) Construct an end-to-end closed-loop process of "perception-reasoning-knowledge generation-decision issuance-verification optimization-memory retrieval" to form an automated knowledge discovery and dissemination mechanism. The network can autonomously perceive new scenarios and adapt quickly, ensuring continuous optimization and evolution under multi-domain integration and large-scale deployment, and realizing intelligent management of network intelligence.

[0065] (3) Integrating meta-learning, reinforcement learning, federated learning and knowledge distillation techniques, federated learning is used to train in parallel on distributed nodes, protecting data privacy while gathering multi-scenario operation and maintenance experience; meta-learning is used to achieve rapid adaptation of small samples under new topology or new scenario; combined with reinforcement learning to dynamically optimize decision-making strategies, and finally master network configuration and traffic scheduling strategies from scratch.

[0066] (4) Through cross-domain knowledge collaboration, network closed-loop management and multi-paradigm learning mechanisms, the network can achieve autonomous learning, dynamic optimization and continuous evolution. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the structure of a cross-domain knowledge-driven intelligent scheduling system for 6G according to the present invention. Detailed Implementation

[0068] To address the challenges of heterogeneity across multiple scenarios in 6G networks, dynamic service demands, and insufficient generalization capabilities of existing automated operation and maintenance models, this invention proposes a cross-domain knowledge collaborative evolution mechanism. This mechanism decomposes network state into an environmental domain (electromagnetic / meteorological), a network domain (topology / resources), and a user behavior domain (service characteristics). It achieves online incremental fusion of heterogeneous data through graph neural networks and attention mechanisms, and dynamically updates the global knowledge base using knowledge distillation technology, enabling collaborative evolution of knowledge across the three domains. Furthermore, a multi-layered closed-loop management architecture is constructed, integrating an end-to-end process of perception → reasoning → decision-making → verification → feedback. This architecture incorporates federated learning, deep reinforcement learning, and meta-learning to achieve cross-scenario policy transfer and continuous optimization.

[0069] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0070] Combination Figure 1 This invention provides a cross-domain knowledge-driven intelligent network scheduling system for 6G, including a cross-domain knowledge collaborative evolution module, a network closed-loop management module, and a multi-paradigm learning decision module;

[0071] The cross-domain knowledge collaborative evolution module is used to decompose the network state into the environment domain, network domain and user behavior domain, realize the online incremental fusion of knowledge in the three domains through graph neural network, and dynamically update the global knowledge base based on knowledge distillation.

[0072] The network closed-loop management module is used to construct a functional closed loop of perception-reasoning-knowledge generation-decision issuance-verification optimization-memory retrieval;

[0073] The multi-paradigm learning decision module is used to integrate meta-learning, reinforcement learning, federated learning, and self-supervised learning to achieve rapid policy transfer and continuous evolution.

[0074] As a specific example, the cross-domain knowledge collaborative evolution module includes a domain splitting unit, an online fusion unit, and a knowledge base updating unit;

[0075] The domain splitting unit is used to define environmental domain data as meteorological parameters, electromagnetic characteristics and geographic information, network domain data as topology, resource status and performance indicators, and user behavior domain data as business type, service requirements and task priority.

[0076] The online fusion unit uses a graph neural network to uniformly encode the heterogeneous data across the three domains, and dynamically allocates weights for each domain through an attention mechanism. These attention weights are calculated using the Softmax function.

[0077]

[0078] in, For domain Attention weights , For domain ,domain eigenvectors, To adjust the parameters;

[0079] The knowledge base update unit compresses new and old knowledge through knowledge distillation, and defines a distillation loss function. :

[0080]

[0081] in, For cross-entropy loss, Let KL divergence be the KL divergence. This is the balance coefficient.

[0082] As a specific example, the environmental domain data includes air density, wind speed, rainfall, channel quality, interference source distribution, coverage blind spots, and path attenuation; the network domain data includes multi-hop topology, air-to-ground handover delay, network slice resource allocation, link utilization, SNR, BER, and latency jitter; and the user behavior domain data includes traffic characteristics, flight paths, service level requirements, and task priorities.

[0083] As a specific example, the graph neural network employs a knowledge graph embedding method based on TransE, using knowledge graph triples. Embedded target function for:

[0084]

[0085] in, , , These are head entities, relations, and tail entities, where an entity refers to the subject in the knowledge graph, a relation refers to the semantic relationship between entities, and a tail entity refers to the object in the knowledge graph. These are the embedding vectors for the head entity, relation, and tail entity, respectively. For the set of positive sample triples, For negative sample set, This is a marginal parameter used to control the spacing between positive and negative samples.

[0086] As a specific example, the network closed-loop management module includes a multi-domain perception layer, an intent parsing layer, a decision execution layer, and a verification feedback layer;

[0087] The multi-domain perception layer is used to deploy distributed sensors to collect data from three domains and output multimodal feature vectors in a unified format:

[0088]

[0089] in For time steps; The unified eigenvector representing time t; Indicates environmental domain characteristics (such as meteorology, electromagnetic, etc.); Represents network domain characteristics (such as topology, resources, performance, etc.); Indicates user behavior domain characteristics (such as business type, priority, etc.);

[0090] The intent parsing layer uses a lightweight NLP model to convert natural language instructions into structured intents. :

[0091]

[0092] in For maximum delay, For minimum bandwidth, To circumvent the strategy, Prioritize the scene;

[0093] The decision execution layer uses a Graph Convolutional Network (GCN) to propagate cross-domain knowledge and generate a globally optimal policy. The feature update formula for the GCN is as follows:

[0094]

[0095] in, For nodes In the Features of the layer For the set of neighboring nodes, For the first The learnable parameter matrix of the layer, Here, j represents the activation function; j represents the index of the neighboring node.

[0096] The verification feedback layer monitors throughput, packet loss rate, latency fluctuations, and energy consumption in real time, and adjusts the strategy through reinforcement learning feedback. The reward update formula for reinforcement learning is as follows:

[0097]

[0098] in, This represents the time difference error, used for policy updates; For instant rewards, As a discount factor, The current state value function, Estimate the value function for the next state.

[0099] As a specific example, the intent parsing layer also includes an experiential knowledge retrieval subunit, used to calculate the semantic similarity between the current intent vector and the experiential knowledge triples. :

[0100]

[0101] in, This is the current intent vector. For the first A strategy template, when When the value exceeds the threshold, the recommended activation strategy template is provided.

[0102] As a specific example, the verification feedback layer will perform feedback injection into the knowledge graph to form new triples. :

[0103]

[0104] in , For entities, To achieve a closed-loop evolution of intention-experience-knowledge in the relationships between entities.

[0105] As a specific example, the multi-paradigm learning decision module includes federated learning units, meta-learning units, and reinforcement learning units;

[0106] The federated learning unit is used to aggregate the local model parameters of distributed nodes, and the global model parameter update formula is:

[0107]

[0108] in, These are global model parameters. For nodes Local parameters, For nodes The amount of data, Total data volume;

[0109] The meta-learning unit optimizes meta-parameters through inner and outer loop training, with the inner loop fine-tuning task-specific parameters:

[0110]

[0111] Outer loop updates meta parameters:

[0112]

[0113] in , These are the learning rates for the inner and outer loops, respectively. The loss function; For the task Dedicated model parameters; These are the meta-parameters, i.e., the initial shared parameters; For the task In parameters The following losses; For the task Parameters after fine-tuning The following losses; For parameters The gradient; For meta-parameters The gradient;

[0114] The reinforcement learning unit employs a deep reinforcement learning optimization strategy, with the objective function being:

[0115]

[0116] in, As a strategy, For trajectory, For strategy Expected return; For a moment Discount factor; For a moment Instant rewards; In strategy Lower trajectory The expectation.

[0117] As a specific example, the federated learning unit also introduces node performance weights, the formula for which the weights are calculated is:

[0118]

[0119] in, For nodes At any moment Aggregate weights; For nodes Comprehensive performance metrics (such as accuracy, latency, etc.); For other nodes Performance metrics; The total number of nodes participating in federated learning; T is the temperature factor, which controls the sensitivity of the control weights.

[0120] This invention also provides a cross-domain knowledge-driven intelligent scheduling method for 6G networks. This method is based on the aforementioned cross-domain knowledge-driven intelligent scheduling system for 6G networks. The method is applied to low-altitude traffic control scenarios and includes cross-domain networking, intent-guided agent management, and swarm path planning. It optimizes throughput, fault recovery time, traffic prediction accuracy, decision latency, resource fairness, path efficiency, and task success rate.

[0121] Through cross-domain networking, intent-guided intelligent agent management, and swarm path planning verification in low-altitude traffic control scenarios, this invention significantly outperforms existing solutions in terms of throughput, fault recovery speed, traffic prediction accuracy, decision latency, resource fairness, path efficiency, and task success rate. It can effectively support the autonomous learning, autonomous optimization, autonomous management, and autonomous evolution of 6G networks, and is applicable to complex 6G scenarios such as integrated space-ground communication and low-altitude traffic control. It provides a brand-new solution for the intelligent operation and maintenance of 6G networks and has significant theoretical and application value.

[0122] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a cross-domain knowledge-driven intelligent scheduling method program for 6G. When the processor executes the computer program, it implements the steps in the various embodiments of the cross-domain knowledge-driven intelligent scheduling method for 6G described above; alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various system embodiments described above, such as a cross-domain knowledge collaborative evolution module.

[0123] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0124] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0125] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0126] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0127] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0128] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the connection relationship between the modules of this invention indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0129] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A cross-domain knowledge-driven intelligent network scheduling system for 6G, characterized in that, This includes a cross-domain knowledge collaborative evolution module, a network closed-loop management module, and a multi-paradigm learning decision-making module; The cross-domain knowledge collaborative evolution module is used to decompose the network state into the environment domain, network domain and user behavior domain, realize the online incremental fusion of knowledge in the three domains through graph neural network, and dynamically update the global knowledge base based on knowledge distillation. The network closed-loop management module is used to construct a functional closed loop of perception-reasoning-knowledge generation-decision issuance-verification optimization-memory retrieval; The multi-paradigm learning decision module is used to integrate meta-learning, reinforcement learning, federated learning, and self-supervised learning to achieve rapid policy transfer and continuous evolution.

2. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 1, characterized in that, The cross-domain knowledge collaborative evolution module includes a domain splitting unit, an online fusion unit, and a knowledge base updating unit; The domain splitting unit is used to define environmental domain data as meteorological parameters, electromagnetic characteristics and geographic information, network domain data as topology, resource status and performance indicators, and user behavior domain data as business type, service requirements and task priority. The online fusion unit uses a graph neural network to uniformly encode heterogeneous data across three domains, and dynamically allocates weights for each domain through an attention mechanism. These attention weights are calculated using the Softmax function. ； in, For domain Attention weights , For domain ,domain eigenvectors, To adjust the parameters; The knowledge base update unit compresses new and old knowledge through knowledge distillation, and defines a distillation loss function. : ； in, For cross-entropy loss, Let KL divergence be the KL divergence. This is the balance coefficient.

3. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 2, characterized in that, The environmental domain data includes air density, wind speed, rainfall, channel quality, interference source distribution, coverage blind spots, and path attenuation; the network domain data includes multi-hop topology, air-to-ground handover delay, network slice resource allocation, link utilization, SNR, BER, and latency jitter; and the user behavior domain data includes traffic characteristics, flight paths, service level requirements, and task priorities.

4. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 2, characterized in that, The graph neural network employs a knowledge graph embedding method based on TransE, using knowledge graph triples. Embedded target function for: ； in, , , These are head entities, relations, and tail entities, where an entity refers to the subject in the knowledge graph, a relation refers to the semantic relationship between entities, and a tail entity refers to the object in the knowledge graph. These are the embedding vectors for the head entity, relation, and tail entity, respectively. For the set of positive sample triples, For negative sample set, This is a marginal parameter used to control the spacing between positive and negative samples.

5. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 1, characterized in that, The network closed-loop management module includes a multi-domain perception layer, an intent parsing layer, a decision execution layer, and a verification feedback layer; The multi-domain perception layer is used to deploy distributed sensors to collect data from three domains and output multimodal feature vectors in a unified format: ； in For time steps; The unified eigenvector representing time t; Indicates the characteristics of the environmental domain; Represents network domain characteristics; Represents user behavior domain characteristics; The intent parsing layer uses a lightweight NLP model to convert natural language instructions into structured intents. : ； in For maximum delay, For minimum bandwidth, To circumvent the strategy, Prioritize scenarios; The decision execution layer uses a Graph Convolutional Network (GCN) to propagate cross-domain knowledge and generate a globally optimal policy. The feature update formula for the GCN is as follows: ； in, For nodes In the Features of the layer For the set of neighboring nodes, For the first The learnable parameter matrix of the layer, Here, j represents the activation function; j represents the index of the neighboring node. The verification feedback layer monitors throughput, packet loss rate, latency fluctuations, and energy consumption in real time, and adjusts the strategy through reinforcement learning feedback. The reward update formula for reinforcement learning is as follows: ； in, This represents the time difference error, used for policy updates; For instant rewards, As a discount factor, The current state value function, Estimate the value function for the next state.

6. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 5, characterized in that, The intent parsing layer also includes an empirical knowledge retrieval subunit, used to calculate the semantic similarity between the current intent vector and the empirical knowledge triples. : ； in, For the current intent vector, For the first A strategy template, when When the value exceeds the threshold, the recommended activation strategy template is provided.

7. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 5, characterized in that, The verification feedback layer will perform feedback injection into the knowledge graph to form new triples. : ； in , For entities, To achieve a closed-loop evolution of intention-experience-knowledge in the relationships between entities.

8. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 1, characterized in that, The multi-paradigm learning decision module includes federated learning units, meta-learning units, and reinforcement learning units; The federated learning unit is used to aggregate the local model parameters of distributed nodes, and the global model parameter update formula is: ； in, These are global model parameters. For nodes Local parameters, For nodes The amount of data, Total data volume; The meta-learning unit optimizes meta-parameters through inner and outer loop training, with the inner loop fine-tuning task-specific parameters: ； Outer loop updates meta parameters: ； in , These are the learning rates for the inner and outer loops, respectively. The loss function; For the task Dedicated model parameters; These are the meta-parameters, i.e., the initial shared parameters; For the task In parameters The following losses; For the task Parameters after fine-tuning The following losses; For parameters The gradient; For meta-parameters The gradient; The reinforcement learning unit employs a deep reinforcement learning optimization strategy, with the objective function being: ； in, As a strategy, For trajectory, For strategy Expected return; For a moment Discount factor; For a moment Instant rewards; In strategy Lower trajectory The expectation.

9. The cross-domain knowledge-driven intelligent network scheduling system for 6G as described in claim 8, characterized in that, The federated learning unit also introduces node performance weights, the formula for which the weights are calculated is: ； in, For nodes At any moment Aggregate weights; For nodes Comprehensive performance indicators; For other nodes Performance metrics; The total number of nodes participating in federated learning; T is the temperature factor, which controls the sensitivity of the control weights.

10. A cross-domain knowledge-driven intelligent network scheduling method for 6G, characterized in that, This method is based on the 6G-oriented cross-domain knowledge-driven network intelligent scheduling system according to any one of claims 1 to 9. The method is applied to low-altitude traffic control scenarios, including cross-domain networking, intent-guided intelligent agent management and swarm path planning, and optimizes the indicators of throughput, fault recovery time, traffic prediction accuracy, decision delay, resource fairness, path efficiency and task success rate.

Images